Write the AI Clause So a Detector Report Is Evidence, Not the Finding

An institution-level clause has 4 questions to answer, and a syllabus cannot answer them for the institution. What counts as unauthorized AI use across every course, and not only in one. Whether a detector report is admissible at a hearing. What has to sit beside that report before a charge is filed. Which office decides, and on what grounds a student can appeal. A teaching centre writes guidance. A code writes the rule a hearing panel is bound by.

The City University of New York settled scope without inventing a new offence. Its definition of cheating already covered unauthorized use of material, information, notes, study aids and devices during an academic exercise. Artificial intelligence systems were added to that list in 2024 ¹. The same policy treats AI-generated content used without citing the AI as plagiarism, whether the text was submitted whole or paraphrased ¹.

The advantage there is procedural. Charge a case under the existing cheating rule and it inherits the notice, hearing and appeal machinery already in use. A standalone AI offence written outside that section risks leaving its procedure to be worked out in the first contested case.

This page is about a university code. A K-12 district writing the same language works under different constraints, and the district policy template with a detector clause covers those.

Because those who make and study detectors say it cannot: a vendor, a teaching centre and a peer-reviewed study each stop short of calling a report proof. Turnitin said in 2023, when it launched the feature, that it makes no determination of misconduct even on text similarity, and that the decision rests with the educator and institutional policy ². The University of Pittsburgh's teaching centre puts it as a limit on evidence: detectors are not reliable enough to act as proof ³.

Human's documentation describes its own output in similar terms, which a drafter can borrow. Its reading is an estimate of how much of a paper reads as machine-written or machine-edited, sentence by sentence, with a document verdict of Human, Mixed or AI, and what it never reports is a probability that a person cheated. Human says it this way: "This is an estimate from our detector. Treat a flag as a reason to look closer, not as a finding." A code whose evidence sentence says no more than that has said enough.

The arithmetic is worth doing. Vanderbilt set the 1 percent false positive rate Turnitin claimed against its own volume: 75,000 papers went to Turnitin in 2022. If the vendor's rate carried over to Vanderbilt's papers, all of them human-written, around 750 papers could have been incorrectly labelled as partly AI-written inside a single year ⁴. A rate that reads as tolerable on a vendor page reads differently as a count of papers an integrity office might have to review.

Independent work arrives at the same place from another direction. A 2024 study of detectors under adversarial editing, published through the International Journal of Educational Technology in Higher Education, concluded that the tools cannot currently be recommended for determining whether violations of academic integrity have occurred ⁵.

Banning the tool does not follow from any of it, and a code should not try to settle procurement. Pittsburgh's teaching centre supports no AI detection tool at all ³. A code that keeps one running can still refuse to let a score close a case.

Both choices survive a clause that describes what a report weighs. A clause naming a product fits only one of them.

The narrow version of the question, whether a score alone can support a finding, is better settled in the code than improvised at a hearing.

The clause should name the corroborating material itself, so a panel is not inventing a standard case by case: a documented conversation with the student before any charge is filed, process evidence such as draft history or version records, or a comparison against work the same student produced under supervision. At least one of those sits beside any detector report before a charge proceeds.

CUNY writes the conversation into procedure, where good practice alone would leave it optional. A faculty member who suspects a violation reviews the facts and circumstances of the suspected violation with the student, whenever that is feasible ¹.

That conversation can do 2 jobs: give the student an early chance to explain, and, if the code also requires a written record, leave a dated account of what was said, which is what a file lacks weeks later at a hearing.

Process evidence needs a caution written in beside it. A 2026 framework in Frontiers in Artificial Intelligence warns that process documentation should be read as evidence of engagement rather than proof of independent authorship, since process artifacts are not immune to fabrication ⁶. Draft history corroborates a case without settling it. A clause that treats a document timeline as decisive has moved the same error to a friendlier exhibit.

The cheapest corroboration is the kind that exists before anyone is suspicious. A course that requires an AI use statement on every assignment has the student's own account of the tools on file at submission, written when nothing was at stake. What standard that account is then weighed against is a separate decision, and the burden of proof an AI misconduct hearing runs on is where an integrity office settles it.

A body, not the person who noticed. Ohio State routes alleged academic misconduct to a formal hearing before the Committee on Academic Misconduct, a standing committee of the University Senate drawn from faculty, graduate students and undergraduates ⁷. Separating the accuser from the decider is a structural protection an AI clause can build in, and it takes one sentence to write.

What an accused student is owed belongs in the clause itself, since a cross-reference to a manual nobody opens goes unread. Where a student denies the allegation, CUNY's procedures set a floor of 3 things: written notice of the charges, the right to appear before the committee, and the right to present witness statements or call witnesses ¹.

Appeal grounds carry more weight in an AI case than in an ordinary one, because the accusation often rests on an instrument the student cannot inspect. The University of Texas at Austin limits a conduct appeal to 3 grounds: significant procedural error, new information that was not reasonably foreseeable, and a sanction significantly disproportionate to the violation ⁸.

Under a rule like that one, draft history the student already held may not count as new information, which is a reason to get it on the record before the hearing. And a failure to disclose the tool is a procedural error only where the procedure requires disclosure, which is the reason to write the requirement in.

One route is easy to leave out of a draft. At some institutions an instructor can lower a grade without filing, and then no hearing takes place. Purdue's 2025 guide points a student whose grade was reduced unfairly for alleged academic dishonesty at the grade appeals system ⁹. A code describing only the formal path leaves that case unreviewed. How an AI detection appeals process is designed is the companion decision.

4 parts carry the clause, and each answers a question a hearing panel will otherwise answer on the spot. Scope, evidence, corroboration, decider. The draft below is written for a university code rather than a syllabus. It carries no vendor name and no threshold, and it should go past general counsel before it reaches a faculty senate.

Scope. Unauthorized use of a generative artificial intelligence system in preparing submitted work is cheating under this section, whether the generated text appears whole, in part or paraphrased, and whether or not a specific tool is named in the course syllabus.

Evidence. A report from an AI writing detector is investigative material. It is not a finding of responsibility and may not, on its own, support one.

Corroboration. No charge under this section proceeds on a detector report alone. The record must also contain a documented discussion with the student, process evidence of how the work developed, or a comparison with work the student produced under supervision.

Decider. Responsibility is determined by the hearing body named in this code and not by the instructor who raised the suspicion. The notice of charge discloses any detection tool used and the setting at which it was run. A grade reduced in place of a charge is reviewable through the grade appeal process.

Two things stay out of the code. A numeric threshold belongs in operational guidance an integrity office can revise without a senate vote, and vendors do not publish a common scale, so a percentage from one tool cannot be assumed to mean what it means on another. Product names stay out for the reason in reverse: procurement moves faster than a code, and a clause naming a supplier is out of date the season the contract changes.